LG-TOM: Language Grounded Theory of Mind Modeling for Multi-agent Communication

Huao Li, Michael Lewis, Katia P. Sycara

Keywords: emergent communication, multi-agent reinforcement learning, social reasoning, large language models

Abstract: Emergent communication refers to the process by which multiple agents learn to develop efficient protocols for sharing information in collaborative tasks. Agents typically learn through interaction with the environment, using reinforcement learning to optimize protocols for task completion. However, the sparse task rewards can lead to unstable training and poor generalization, especially in partially observable environments and decentralized training setups. To address these challenges, we propose \textbf{LG-TOM}, a novel \textcolor{blue}{online RL framework} that enables agents to learn from social interactions via \textbf{Theory of Mind (ToM) modeling with language grounding}. Specifically, we design a belief estimation network that leverages priors from large language models (LLMs), allowing agents to reason about how their communication influences the belief states of others. We then compute social influence as an intrinsic motivation reward, encouraging agents to share information that positively impacts teammates. Experimental results demonstrate that LG-TOM improves communication effectiveness over baselines in multi-agent collaborative tasks \textcolor{blue}{including complex social dilemmas.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 8261